The construction of China’s largest accelerator cluster represents a crucial step in Beijing’s quest to develop a digital infrastructure independent of Nvidia’s hardware monopoly.
Article Index:
- Largest computing cluster for Huawei chips in Mongolia
- Allocating processors for inference tasks instead of training
- Supply chain hurdles and advanced memory shortages
- Comprehensive transformation in China’s tech sector
- Huang’s warnings regarding the erosion of CUDA ecosystem dominance
- FAQs
Largest computing cluster for Huawei chips in Mongolia
Chinese startup “DeepSeek” is preparing to deploy a massive computing cluster comprising at least 160,000 new-generation “Ascend 950 DT” artificial intelligence processor chips from “Huawei Technologies” inside a giant gigawatt-scale data center in the city of “Ulanqab” in the “Inner Mongolia” region, according to Bloomberg. This initiative constitutes one of the largest globally known computing clusters relying entirely on Chinese-made AI accelerators.
This step marks a milestone in China’s tireless efforts to build a digital infrastructure ecosystem completely independent of American “Nvidia” hardware, whose advanced processors were barred from sale to Chinese companies by strict US export controls, forcing domestic developers to innovate resilient national alternatives.
Allocating processors for inference tasks instead of training
DeepSeek intends to use Huawei chips to perform inference tasks—computational operations dedicated to running models and delivering answers to end users on a large scale—rather than using them for primary initial training operations, even though Huawei designed the Ascend processor to handle both computational burdens efficiently. To this moment, DeepSeek continues to rely on its available Nvidia accelerators to accomplish the arduous training task that requires exceptional computational capacity and architecture.
This strategy reflects a practical approach to dividing technical burdens; using domestic chips in daily inference services provides a low-cost advantage and the freedom to expand without fear of sudden Western supply chain disruptions.
Supply chain hurdles and advanced memory shortages
Nevertheless, estimates indicate that delivering the full shipment of 160,000 chips may take more than a full year due to complex production capacity constraints and supply chain bottlenecks. Severe shortages in micro-components, especially ultra-fast memory chips, cap Huawei’s production of Ascend processors at no more than a few hundred thousand units this year.
Huawei also finds itself forced to balance mounting demands from major Chinese companies with attempts to export limited quantities abroad. In this context, DeepSeek—seeking quantities exceeding what it requested—has asked the Chinese government in Beijing to pressure Huawei to grant it top priority to accelerate delivery speeds, while the two Chinese companies remained silent and refrained from official comment.
Comprehensive transformation in China’s tech sector
This record order reflects a wide-ranging structural transformation underway in China’s artificial intelligence industry, as Beijing urges national research labs and companies to completely abandon foreign hardware. Many companies have already begun adapting and tuning their algorithms to run seamlessly with Huawei processors. DeepSeek’s fourth model was optimized to run on Ascend processors upon its release in April, prompting Chinese cloud computing providers to order hundreds of thousands of chips.
This planned cluster will leap forward by a massive margin compared to previous projects; just six months ago, the city of Shenzhen announced the launch of its first Ascend processor cluster comprising only 10,000 chips. DeepSeek, which raised over $7 billion in June to finance its infrastructure, relies on the Inner Mongolia facility as a cornerstone for its future projects.
Huang’s warnings regarding the erosion of CUDA ecosystem dominance
Nvidia CEO Jensen Huang previously described running DeepSeek models on Huawei chips as a terrifying outcome for the United States. Huang explained that export restrictions are pushing the Chinese software ecosystem toward adopting domestic hardware and gradually moving away from Nvidia’s proprietary “CUDA” software platform.
Every model optimized to run on Ascend chips strengthens an integrated software environment whose need for Nvidia decreases over time—a profound strategic shift that, if settled and rooted, will become impossible for American companies to reverse or regain their previous dominance over in the future.
FAQs
Q: What is the size of DeepSeek’s order for new Huawei processors?
A: The company plans to deploy at least 160,000 Ascend 950 DT chips in Inner Mongolia.
Q: What tasks will Huawei chips handle at the DeepSeek facility?
A: The chips will be dedicated to inference tasks and running models for users, while training remains on other accelerators.
Q: Why does this step worry American company Nvidia?
A: Because it contributes to building an independent software environment that moves entirely away from the CUDA ecosystem, undermining the dominance of US technologies.
